hesela.devPrecision
Precision is the fraction of positive predictions that match observed positive outcomes at a specified decision threshold.
For binary failure prediction, precision = TP / (TP + FP), where TP counts correct positive predictions and FP counts false positives. Its denominator is alerts, not healthy observations. It is undefined when no positives are predicted; report any software convention used for that case. [1] A different failure prevalence can change precision even when recall and false positive rate stay fixed. [2]
Relations
- related
- recall, false-positive-rate, class-prevalence, average-precision, smart
Limitations
- Specify the failure label, prediction horizon and counting unit; a drive-day alert is not automatically a distinct drive selected for replacement. [3]
Citations
- scikit-learn: precision_score; binary precision and zero-division behavior.
- Saito and Rehmsmeier. The Precision-Recall Plot Is More Informative than the ROC Plot When Evaluating Binary Classifiers on Imbalanced Datasets. PLOS ONE 10(3): e0118432 (2015).
- Hesela: disk failure prediction, metric denominators and a synthetic alert-count example.